Abstract: Based on UAV multispectral imageries combined with field-measured samples from typical no-till reseeded grasslands in alpine regions, four models, including multiple linear regression (MLR), ElasticNet, random forest (RF), and support vector regression (SVR), were developed to compare the performance of aboveground biomass (AGB) estimation under different model types (linear and nonlinear models) and data organization strategies (classification modeling by grassland type, overall sample modeling, and overall sample modeling with grassland-type dummy variables). The results showed that sample heterogeneity was evident among different grassland types, and that the relationship between fractional vegetation cover (FVC) and AGB varied with grassland type. In the classification modeling, there was no single optimal model for all grassland types. For each grassland type, ElasticNet performed best for alpine meadow grassland (test R²=0.808), whereas MLR was optimal for alpine desert grassland (R²=0.728). The performance of all models was generally poor for alpine grassland (R₂<0.40). For the overall sample modeling, the test R² values of all models ranged from 0.519 to 0.573. Although SVR showed better performance (R²=0.573), the overall results were inferior to those of classification modeling. The performance of all models did not improve after introducing dummy variables for grassland type. Under the classification modeling strategy, linear models were generally more stable than nonlinear ones. In contrast, under the overall sample modeling strategy, the performance gap between the two types of models narrowed, with SVR showing a relative advantage. Variable contribution analysis indicated that the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and FVC were the main explanatory variables for AGB estimation in the overall sample set, though key variables differed among grassland types. These findings demonstrate that sample heterogeneity and data organization strategy are key factors affecting the accuracy of AGB estimation for no-till reseeded grasslands in alpine regions. The classification modeling is more targeted and stable than overall modeling and dummy-variable treatment.

Key words: alpine grassland, grassland remote sensing, multispectral imagery, vegetation index, regression model